Qwythos-9B Claude Mythos GGUF: 1M‑token, multimodal, uncensored reasoning model

The **Qwythos-9B-Claude-Mythos-5-1M-GGUF** model, released by Empero AI, is a GGUF‑quantized version of the base Qwythos‑9B Claude Mythos model. It is built on the Qwen3.5‑9B architecture and has been post‑trained on over 500 million tokens of high‑quality Claude Mythos and Claude Fable traces, giving it strong reasoning abilities (evidenced by +34 pts MMLU, +30 pts GSM8K‑strict, and +19 pts GSM8K‑flex over the base). The model supports native function calling per the Qwen3.5 specification and ships with a 1,048,576‑token (1 M) context window enabled by YaRN rope‑scaling.

The repository provides multiple quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0, BF16) as well as MTP‑enabled variants for llama.cpp draft speculation. The recommended default is the Q4_K_M quant for a good quality‑size trade‑off. A separate CLIP‑style vision projector (mmproj‑F16.gguf) allows image input, inheriting the vision capabilities of the original Qwen3.5‑9B model such as detailed description, OCR, chart reading, and basic spatial reasoning. Vision functionality works with any of the text GGUF files when paired with the projector.

Qwythos is marketed as an uncensored, agentic model suitable for demanding domains like cybersecurity, biomedical research, and general tool‑use. It emits a `<think>...</think>` block at the start of each response to expose its chain‑of‑thought reasoning, and the README recommends sampling settings (temperature 0.6, top‑p 0.95, top‑k 20, repeat‑penalty 1.05) to avoid repetition loops. The model can be run with llama.cpp, Ollama, LM Studio, Jan, KoboldCpp, or any GGUF‑compatible runtime, and it includes support for OpenAI‑compatible vision APIs when served via llama‑server.

All weights are released under the Apache‑2.0 license, inherited from the Qwen3.5‑9B base model, and the project encourages users to add their own safety layers due to the model's uncensored nature.